Inferring left-turn information from mobile crowdsensing

Abstract

Left turns are known to be one of the most dangerous driving maneuvers. An effective way to mitigate this safety risk is to install a left-turn enforcement—for example, a protected left-turn signal or all-way stop signs—at every turn that preserves a traffic phase exclusively for left turns. Although this protection scheme can significantly increase the driving safety, information on whether or not a road segment (e.g., intersection) has such a setting is not yet available to the public and navigation systems. This disclosure presents a system that exploits mobile crowdsensing and deep learning to classify the protection settings of left turns.

What is claimed is: 
     
       1. A method for classifying left turns of a vehicle, comprising:
 receiving, by a processor of a mobile device, a sensor signal from a sensor residing in the mobile device while the vehicle is moving, where the signal is indicative of angular speed of the vehicle 
 constructing, by the processor, a feature vector from data received from the sensor; 
 training, by the processor, a classifier in part based on the feature vector, where the classifier outputs a probability that a left turn made by the vehicle is a protected left turn; 
 identifying, by the processor, one or more bumps in the sensor signal; and 
 classifying, by the processor, a left turn made by the vehicle as being a protected left turn based on number of bumps identified in the sensor signal, where the protected left turn has a traffic indicia which causes a time period assigned exclusively to making a left turn. 
 
     
     
       2. The method of  claim 1  further comprises classifying a left turn made by the vehicle as being a protected left turn in response to identifying one bump in the sensor signal, classifying a left turn made by the vehicle as being an unprotected left turn in response to identifying two bumps in the sensor signal. 
     
     
       3. The method of  claim 1  further comprises
 detecting, by the processor, a left turn made by the vehicle using signals from one or more sensors residing in the mobile device; and 
 identifying the one or more bumps in the signal in response to detecting a left turn of the vehicle. 
 
     
     
       4. The method of  claim 1  further comprises:
 converting, by the processor, the sensor signal from a coordinate system associated with the mobile phone to a geographic coordinate system, where the sensor signal is report by the sensor relative to the coordinate system associated with the mobile phone; and 
 deriving, by the processor, an angular speed signal from the converted sensor signal, where the angular speed signal is indicative of angular speed of the vehicle about a vertical axis of the geographic coordinate system. 
 
     
     
       5. The method of  claim 1  further comprises
 receiving, by the processor, geolocation of the mobile device; 
 associating, by the processor, geolocation with data points of the sensor signal, where the sensor signal is expressed as a time series; and 
 determining, by the processor, two bumps in the sensor signal to be part of the same left turn based on proximity of geolocation of the two bumps to each other. 
 
     
     
       6. The method of  claim 5  further comprises receiving geolocation of the mobile device as readings of a global positioning system (GPS) residing in the mobile device. 
     
     
       7. The method of  claim 1  further comprises measuring the angular speed of the vehicle using a gyroscope residing in the mobile device. 
     
     
       8. The method of  claim 1  wherein the classifier is further defined as a neural network. 
     
     
       9. A method for classifying left turns made by a vehicle, comprising:
 receiving, by a processor, a plurality of feature vectors, where each feature vector is indicative of a left turn made by a vehicle and each feature vector includes geolocation of the vehicle during the left turn and time series data for the angular speed of the vehicle during the left turn; 
 spatially clustering, by the processor, the plurality of feature vectors based on geolocation to form one or more clusters of feature vectors; 
 assigning, by the processor, a ground truth to the one or more clusters; and 
 training, by the processor, a classifier with feature vectors in the one or more clusters using supervised learning, where the classifier outputs a probability that a left turn made by a vehicle is a protected left turn and a protected left turn has traffic indicia which creates a time period assigned exclusively to making a left turn. 
 
     
     
       10. The method of  claim 9  further comprises
 collecting sensor data from one or more sensors residing in a mobile device while a vehicle is making a left turn; 
 constructing a feature vector from the sensor data associated with each left turn; and 
 uploading the feature vectors from the mobile device to a server. 
 
     
     
       11. The method of  claim 9  wherein each feature vector further includes time series data from an accelerometer in the vehicle during the left turn. 
     
     
       12. The method of  claim 9  further comprises, for each cluster, concatenating time series data from the feature vectors belong to a given cluster to form training data for the classifier, and randomly permutating sequence of time series data from the feature vectors belong to the given cluster to form additional training data. 
     
     
       13. The method of  claim 9  further comprises spatially clustering the plurality of feature vectors using a density-based spatial clustering of applications with noise method. 
     
     
       14. The method of  claim 9  wherein the classifier is further defined as a recurrent neural network. 
     
     
       15. The method of  claim 9  further comprises
 receiving a set of feature vectors indicative of a left turn made by a vehicle at a particular geolocation; and 
 classifying the left turn as one of a protected left turn or an unprotected left turn using the trained classifier. 
 
     
     
       16. The method of  claim 15  further comprises selecting a route for a vehicle based in part on the classification of the left turn at the particular geolocation. 
     
     
       17. The method of  claim 1  further comprises updating a database with an indicator that the left turn made by the vehicle is a protected left turn.